Suggested fix for issue #4706, including enhancements to the handling of #[rustfmt::skip].

#4706 issue was caused because format_impl() did not copy the skipped range from the local context into the visitor.context. The same issue was corrected also in format_trait() and rewrite_macro_with_items().

While testing I found that there are cases where self.line_number in push_skipped_with_span() is incorrect (usually 0 or 1). That caused an incorrect skipped range to be created. Therefore, I had to make changes to this function. The main change is that the beginning of the skipped range mainly depends on the beginning of the attributes block, instead of self.line_number.

Another major change to push_skipped_with_span() is related to the following comment:

This comment is not correct in general, as in several cases only the first attribute in an attributes block is not skipped (this was also partly discussed in PR #4707). Therefore, my change to push_skipped_with_span() assumes that only the first attribute line should not be skipped. If this is the correct approach, then the above comment should be removed. Otherwise, push_skipped_with_span() should be changed so the beginning of the skipped range will depend on the end of the attributes block, instead of its beginning.

I am aware that the above changes are more general than the specific #4706 issue, but there doesn't seem to be a good way to fix only this issue, because of the required change for push_skipped_with_span() to handle wrong self.line_number that impacts other skip cases.

Because of the general changes, the test cases file include tests for skipping fn/impl/trait/macro. In addition, a test was added in report.rs to make sure that the generated skip_ranges is correct.